Loading summary
Transcript56 lines
- [00:00]
A
Steve, you do a lot of work at the intersection of AI and markets. What does compute demand actually look like right now?
- [00:07]
B
Tight. Right. Demand is extremely high and demand is higher than supply and looks to be higher than the supply in the foreseeable future. And it's been, I think, getting tighter.
- [00:18]
A
Okay, so is it as easy to say because compute demand is tight, therefore no AI bubble?
- [00:27]
B
Well, I think bubble is a different category. Bubble has to do with asset prices. Right? You know, if a stock sort of triples in a month and comes down, you know, by 20%, is that a bubble? I don't know. I mean, so asset prices can get ahead of fundamentals. Ultimately. I think bubble is kind of a challenging thing. I've been saying since 2023. This was. May I post this on Twitter that there was a baby bubble in AI? I think people were talking about baby bubble. I said, great, I love a baby bubble. Baby eventually grows up. So I mean, to me a bubble means a different thing. It means an entire trajectory. And we are certainly part of that trajectory. Eventually we're going to overbuild, but we're not there yet.
- [01:07]
A
Okay, so you have this data from your company, Silicon Data, the LLM Token Expenditure Index. And I'm looking at this right now, it says essentially AI users willingness to pay for tokens is down about 20% since maybe what does that mean? Is that token prices are coming down or that demand for tokens is coming down?
- [01:30]
B
Actually, it's kind of neither. It's a little subtle. And that's why we wrote a little social media post which led to this index getting picked up by, I think, several securities, wrote a little note on it and every other news agency started picking up on it and now has become almost unintentionally by us, this kind of AI bearish index. I actually love this opportunity. To clarify a little bit, what it is is actually a little bit like in the macroeconomics where you have infl measured by expenditure Weighted Price index, the PCE way of measuring inflation. In other words, I'm going to allow consumers to rationally substitute between alternatives they can consume. And instead of fixing a basket and say 20% this, 20% that and whatever, I'm going to allow them to choose however much they want to consume and see what the net price expenditure is and that captures the true inflation. So in this case, we're going to allow for the fact that people can choose different AI models, especially when in an agentic kind of API call context. Now that you have a lot of these platforms that have routing mechanism you can actually choose. Some models are very powerful, but they are quite expensive, right? But some models are open source and much cheaper, but of course they are less powerful. Any given moment people can choose and vote with their feet, so to speak. And when you do that, when you sort of take a look at all the models that are out there and you do an aggregation of the input output tokens specifically, because by the way, like AI, when you ask it a question, does input number of tokens and it comes out announces an output number of tokens, they have different prices, you have to adjust to get like a blend price for the model. And if you then aggregate all the models based on usage of how much people are actually spending, if you look at a million tokens, let's say just normalize to a million on any given day, if people are spending most of the tokens, spending most of their usage on the most frontier closed source proprietary models, whether it's OpenAI or anthropic Claude model, whether it was Opus or Fable or Mythos, those models are going to be much more expensive than the cutting edge open weights models coming from China or elsewhere. And by a factor of maybe 10, right? So once a model comes out, actually the token price doesn't change all that much. What ends up changing is actually every time you have a new model that comes out, the model token is going to have a different price. So that even though ChatGPT4 today, the token price is probably negligible, nobody's using it anymore. But you know, if we look at expenditure, it's going to automatically capture by relevance. And this curve, this chart, this index, when it moves around is mostly moving around by usage. Are people spending most of the tokens on the most expensive models? Are they buying mostly Ferraris or are they mostly buying Corolla? Both cars have a market, both of them get to a place, but people have different preferences, a different usage. So if you look at that chart over the last year since we started tracking this was in December. There was a run up from December through early January and people were using a lot of the advanced OPUS models. I remember the OPUS models really kicked off this whole agentic AI API core programmatic AI usage revolution. And then it came down a little bit. People started using some of the newer open based models that came out. There were some advanced ones that came out, but then in March we saw another wave up and this was the period of the cloud maxing and you saw all these headlines of oh, there's even a competition within Amazon of see who can use the most number of tokens. All the companies were bragging about just how many tokens they were burning. I think Jensen was on some all in podcast interview saying oh if I have this engineer, he better be spending $150,000 worth of tokens otherwise I'd be very worried. So there was that period, right? And then towards end of May, if you track our index, I wrote a little post on Twitter saying oh, it looks as though this index has plateaued a little bit. Maybe people are having second thoughts. Coming to the end of the quarter, the bell come in and realize oh w this is actually a lot more expensive than what we previously thought. So we said perhaps there's going to be some inflection at which point this thing could go either much higher because people just see tremendous productivity gains and roi or maybe this thing could go back down as people become more rational and try to economize on the model choice. And indeed that's what we have seen is just this index has come down. And during this period was also when we heard all the news about how the major companies, tech companies or Uber and saying oh by the way like token expenditure, we've already burned like half year worth of expenditure in a month and this may or may not have led to the type of gains productivity increase that we had expected. In any event, we're going to become more rational about how we maybe we don't field or maybe route every single question through the most powerful model, right? If you try to get the Italian cooking recipe, maybe a less powerful model is sufficient. And so the question is, are people going to become price sensitive and rational about more the quality and cost trade offs. And I think the answer is yes.
- [06:40]
A
So okay, if we're down 20% since May, if I'm sitting here as a individual investor and I look at this chart, what does that actually mean is that, you know, I could see it being bullish. Maybe people are getting smarter about how they use AI, which could be long term, good for the companies maybe or bearish. Because if people are becoming less and less willing to pay for AI tokens, maybe that's bad for the companies. How do you see this?
- [07:09]
B
So I think there's a time horizon thing, right? So remember there was deep seek, right? Deep seek. This was like early last beginning of last year there was the Hojivon's paradox came out. And what was the core intuition? The core intuition is there is that the proliferation of more efficient AI is not bad for AI, even if in the Short run, it could lead to a bit of a panic because the current paradigm or how things are funded and earnings being assumed might change. So right now, what's happening, the reason this might be seen as potentially bearish is because at present Most of the CapEx is coming from the leading labs, OpenAI Anthropic, contracting the provision of compute and training and using their models. And because they charge top dollar, it is seen that they can afford the. But over the long run, as we see especially now that also enterprises and software companies themselves, the cloud providers, even Microsoft, are voicing sort of skepticism about building entire tech stack on single closed source model that could be taken away by regulation and so on. And in any event, it may not be wise for you to build your entire IP on someone else's closed source model. So there's now a life debate and I think there's going to be a purpose for both. Right. So now what we are looking at could be a proliferation of AI capabilities where you have more enterprises, more sovereign AI, so to speak. Sovereign AI is essentially saying more entities, medium sized, large size entities that may want to host their own compute and train their own models and own their own intelligence. And that intelligence goes beyond just their own data by their own sort of institutional know how, right? And rather than sharing all of that with another proprietary closed source model that could come to eventually replicate them and displace them, they will want to actually own that identity themselves. What that means is that you're going to have a proliferation of AI and proliferation of compute that you can imagine that actually broadens the market and make the market a lot bigger. So instead of OpenAI anthropic being all of the demand for compute, you're going to have them being a smaller fraction but of a much bigger overall market
- [09:35]
A
if that were to unfold. My understanding is that OpenAI and Anthropic are heavily subsidizing token usage right now for their users, is that right?
- [09:47]
B
I wouldn't say heavily subsidizing. There is definitely a sense in which in the beginning when you're trying to grow, right? Especially when you're trying to get people to adopt, you know, the first generation of ChatGPT $20 per month subscription plan was certainly cheaper in order to entice usage. And over time we are seeing a shift away from that, oh you can eat buffet kind of a fixed cost subscription plan, Even if it's $200 max plan or whatever, towards a more like electricity, you know, pay as you go, you know, metering system where you're paying for the tokens as you are consuming it. So maybe you pay a fixed price per comes a certain allocation beyond which you pay per usage or something. Right. So we are evolving away from it. And part of the, you know, what we are seeing in this, you know, trend in the token index trend is exactly that, right. As the price increases, people are going to become more conscious. Just like when, you know in any other context, right. If your health insurance was fully covering everything, you're going to go to the hospital and use the most expensive drugs all the time and get the best care all the time. But if you have to actually pay out of pocket for at least part of the procedure, procedure or part of the medicine, you're going to be more conscious and use substitution. Or what does the benefit versus the cost look like? And how should I optimize with respect to the budget? So that's I think what we are be what we'd be looking at.
- [11:11]
A
Okay, Steve, you have this second chart here about GPU rental rates are up 20 to 36% since December, depending on which chip you have. Blackwell, a 100 Hopper H100. When you look at this chart, what are you taking away from this?
- [11:31]
B
So this is very interesting by the way, to me. I'm also coming into this relatively new, even though I've been tracking the AI trade since 2023. It is a very complex kind of usage. Different chips actually do different things. In actual deployment there is model training and then is inference. So if you look at those three chips, B200 is the most advanced chip that came out most recently. H100 has actually been out for almost three years at this point and 800 maybe as long as like five years. Right. So typically speaking, people would tell you by this point a 100 should be well past this lifespan and you should be not used anymore. But what you're looking at there is that per hour rental rate for all three of them are actually higher today than they were at the beginning of the year. Right. What does that mean? That means that there is actually people are demanding and using the compute power of all three chips. But at any given moment they show slightly different trends. And most recently, what has alerted people, alarmed some people, is that if they focus mainly on the H100 workhorse chip, that's the most widely deployed chip for training. But also increasingly inference is that it came down a little bit on the front of the curve, the on demand kind of price. And that has led people because we are in such a momentum driven trade, this AI Thing has become known by more people now everyone's sniffing for every single new sign to say, oh, has everything changed? And people feel like this is kind of a musical chair kind of a thing. Right. But what we see instead is that if you broaden and look at multiple signals and triangulate a, what's really happening is that you see, at least on the front, on the on demand prices, H100 rental price has come down a little bit based on our data. Indeed. But B200 has actually gone up. Right. And A100 in fact stays flat and it's sort of creeping up ever so slightly. Right. In other words, it's very, very stable. What that tells me is that there's a kind of a barbell thing going on where there's a lot of inference demand. By inference we mean like just querying the model. Right. There's a ton of demand, especially because of agentic AI or programmatic calling of the models that has just absolutely increased the total demand such that even an old chip like a 100 is seeing ample demand. That's keeping the price elevated.
- [14:10]
A
Okay, so I'm going to jump in here. Just to clarify. Does that mean the lowest power chip is seeing the biggest demand?
- [14:19]
B
I wouldn't say it's seeing the biggest demand. Rather there is some substitution on the front, the more powerful, more powerful chips in that if, let's say some people are demanding more model training and they want to sort of shift some of the workload from the H100 chip towards the B200 chip. Right. And you can actually see a slight dip in the short run of H100 towards B200. Meanwhile, if inference demand is very strong, you can see the A100 chip still being very strong, robustly demanded. Does that make sense?
- [14:59]
A
It does make sense and it fits an assumption of mine. I don't know if this is verified on the data. Most people are using AI for very general search inquiries or prompting in GPT. Hey, can you get me directions? Can you build my travel itinerary? Build me a workout plant? Just back and forth, Q and A, they're using it as a new Google. Essentially. My inkling is that most people who use AI, that is their exposure to it, they're not trying to build a tech stack or trying to discover a
- [15:32]
B
new drug or inventory.
- [15:34]
A
So is that what you're describing in the data? The most basic level of inquiries is being called upon more and more
- [15:46]
B
basically. I mean, in other words, we don't know exactly, but it is certainly consistent with that theory of Yours. In other words, we are just at the beginning of proliferation of AI. Most people are not using AI in a very sophisticated way. As you say, AI is really powerful, but most people are using it as kind of a glorified search engine that instead of in the past, you Google something, you have to assemble the answer between different search results in your mind or maybe explicitly now, AI prepares the entire answer in a coherent set of paragraphs for you. And that is the use case for the vast majority of people. And that's likely to be the case for a while yet. And that level of proliferation is still happening. Right? But there is also another degree of proliferation is that there is agentic AI. In other words, there's programmatic usage of AI. So today when you ask an AI query by itself, it's already doing that. It's actually activating multiple agents and gathering answers. So if you pull out you on Twitter, you use Grok, probably like the Elon thing actually makes that agentic use very explicit, right? It says, oh, I'm activating like multiple agents and one of them is going onto a YouTube, one of them is going onto Wikipedia, and so on and so forth and then assembling a final answer for you, right? When you are able to use AI to call AI, it really multiplies the amount of compute demand, right? But each of which is not really super high demand, such that this, this type of use can actually be addressed with maybe not even the most powerful chip.
- [17:26]
A
When I think about who has the pricing power in this AI ecosystem, I look at OpenAI, I look at Anthropic, and then a company like Nvidia who really holds the power there. My assumption is that it's Nvidia because they're controlling the most powerful technology that, that OpenAI and Anthropic are sort of relying on. Is that right?
- [17:49]
B
Yeah. I mean, this is my third appearance. I remember the first time I came on, we spoke about this index that I created at Bloomberg called Pricing Power Index. And I told you that the way I capture pricing power is looking at the stability of gross margin. Right? It's not just about the level, obviously level matters. And what comes to Nvidia, what's remarkable is that they have been charging a 75% margin, gross margin for as long as this thing has been out there, right? So they have tremendous pricing power and with tremendous pricing power, obviously the stock is going to be rewarded and it has been rewarded until recently. Then of course, more recently, I think Nvidia has moved sideways and become some people called A bit of a funding short because people are sort of saying, oh, Nvidia has already had this big run. It's unlikely for Nvidia to double, triple easily. But then everyone's looking for that get rich quick kind of stock. So they're looking for other bottlenecks of picks and shovels, so to speak. But generally speaking, yes. I mean, there is tremendous pricing power coming from the Nvidia, but also memory makers, right? And storage. At present, I would say they probably have even more pricing power. They are earning an 80, 90% gross margin. How sustainable? That remains to be seen. It is inconceivable, I think, to anyone rational to think that a gross margin of 80, 90% could be sustained for many, many years. But, you know, so this is like Micron, SK Hynix, SK Hynix, Samsung, and then the storage side, you've got SanDisk and so on. But ultimately this is all connected in that agentic AI, because it's become so easy to call AI. You're going to inevitably create tons of data, right? And tons of data inevitably requires storage of that data that's generated, as is the long context, because the questions and answers are going to become more complex. You have a longer conversation. So that just means you're going to have more demand for memory. And such is the trade. It's always that you have a thing that's pretty cyclical that's been there for the use of something else. We've spoken about this before in my other indices. And then suddenly there's this demand that was not expected and there was no supply capacity because supply takes a while to come online. There's no supply elasticity. Price has to adjust. Right.
- [20:12]
A
How are you thinking about token economics as far as if the technology gets to the point where it's cheaper and cheaper to use tokens and prompt these models, what does that do to a company like Anthropic or OpenAI, which is, I think, actively trying to figure out how to extract value from their users and charge just enough where they don't scare them away, but that they become profitable companies.
- [20:41]
B
I think they each are solving slightly different problems. When you are at that frontier, your business strategy obviously really is immensely important. Anthropic actually has really healthy margins. I mean, on their servicing of the model, they I think, run something like 70, 80% gross margin. And it's almost like software. It's basically like software. Right. So what is costly and the reason why these companies are net, we don't know, because they're not public yet. But from what you do know they're not money making is because they're still investing heavily to train the next generation of models. It's almost like you build a new company that is in the business of selling widgets, everything. You build a factory, the factory builds out widgets and the widgets are really profitable. It costs a cent to make and you sell it for 10 cents. That's really good gross margin. But you also have to keep building new factories. And when you're building a new factory, that new factory has not sold any widgets is going to seen as a pure cost, right? So when you lump the two together, you're going to be in a net loss situation. If you pause everything today, the current output are all grossly, enormously profitable. The question becomes what next? So at some point the models will get smart enough that maybe saturates the use cases for most people. Most people probably don't need arguably models that are super intelligent above a certain threshold. We don't know where that line is. At which point when you keep investing for the even smarter models, how do you then justify that investment? Who are the actual end users who have such a valuable use out of those models that their total usage gives you enough of a tam, right? Enough of a total addressable market to actually justify the initial outlay?
- [22:46]
A
These are the trillion dollar questions I think quite right. Companies are trying to figure it out, investors are trying to figure it out. Steve, I want to ask you about specific investment ideas and trade ideas you might have right now I want to start at a sector level. I know you're looking at earnings very carefully. What stands out to you right now?
- [23:08]
B
The AI trade obviously is three years old and we are getting a little bit, I think wobbly on it. All sorts of questions are surfacing. There is the big number of question like the capex is so enormous people are asking the ROI question. So we're not even concerned so much about the macro picture as much which obviously could cause its own issues. But the general point is that I think the trade that is quite crowded is known by everybody. So at this point there is actually plenty to do and look at outside of AI. Actually I'm still a firm believer in the AI trade. It's just, I think it's less easy over trade than it was three years ago. You want to be sort of tactical and timing and so on and so forth. But what I do actually like is the fact that the economy overall is actually very sound. We just got numbers this morning. Inflation actually seems cool, right? Therefore, the Fed is not going to be in the panic. They want to go hike the hell of the rates and send the economy into a tailspin. In that environment, I like things that are actually your normal pricing, power company, quality compounders that have actually been punished. Things that are kind of boring. Going back to the call, back to the things that we had discussed from our previous interviews. So like some of the software companies, for example, you know, just goofy. An example like FactSet, right? You know, a competitor of my old employer, you know, that company has been punished severely, but it's actually a financial data company that is adopting AI pretty aggressively and now able to interact with that data a lot more, I think efficiently with AI and you can do a lot more. And this goes back to the question we had earlier, right? If FactSet will be incentivized to own their own tech stack, their own AI infrastructure, instead of building their entire IP on top of another proprietary model, they can build their own. Then they become a more powerful company and the data becomes a lot more interactive. Instead of having the quant scale know how to get in the door, you can actually start playing with their data and do a lot more, even from a less technical background that actually expands their customer base. Right. So I actually think there can actually be AI beneficiaries in the software space. I actually like software here. I like productivity software, you know, HubSpot, you know, just give a few examples.
- [25:43]
A
You know, would you buy the basket? Like would you buy the sector of software right now?
- [25:47]
B
Yeah, I mean, you know, I think, yeah, I mean you can never time exact bottom. But I think if you look at software overall, it looks to me to have bottomed. But obviously different pockets have actually done better than others. You know, cyber obviously hasn't even taken a. Because AI in fact made cyber much more important today. In fact, I think Crosstrek is actually surging some 8, 9% because of the IBM news. IBM said some companies pull back on spending because of the Mythos scare on cyber risks. So every time you have advancements in AI, in fact it makes the cyber risk and the defense cyber risk more valuable. And then everything that charges a toll on web traffic, whether it's a cloudflare or even companies. There was this company called Bandwidth. I had never heard of it, but my friend, but my buddy Citrini, they found this name datadog companies that generally put a toe on the traffic. It seems very fashionable these days to put toes on everything they have been doing well. But I actually think the companies that got punished from A forward looking basis. The enterprise software that was questioned because they previously had this perceived pricing, maybe people were quite wondering whether or not that could get disrupted by AI with agents and undermine their pricing power. I actually think they can actually stick around longer than people think. And especially now that I don't know if you saw the statistics, the companies, that is one person or a couple of people, kind of LLCs, that's actually skyrocketing. AI has made it easier than ever to have small companies. And when you have small companies, you want the enterprise software to offload. Some of those things you don't want to do yourselves. You don't want to be vibe coding your own CRM software. You want to benefit from that economy of scale. So I actually think there's opportunities is that sort of thing. Yeah.
- [27:50]
A
So you are not the first person to tell me they like software at these levels. And I don't know if I would dip my toes into software personally. I think one I don't know it or study it well enough. So I'm not going to try to call the bottom on something. So just to recap, fact set ticker is FDS down about 12% this year. This is a stock you like. I want to ask you about this next stock here, which you sent me. Arthur JJ. Gallagher Ticker AJG, it's an insurance company. Why do you like this stock?
- [28:26]
B
Again, going back to the pricing power concept, right? Ajg if you chart the stock, the stock was like this and then went like that and then now it's coming back.
- [28:34]
A
So all the way up, all the way down.
- [28:35]
B
Not all the way down. It has adjusted down some a little bit. Right. If you chart like Costco, by the way, Costco hasn't had this thing, but Costco is kind of following a similar thing. In other words, these are all companies that are fantastic, perfectly good companies, very stable margins, very tight lock on their markets. The only thing issue was that they were priced more expensively. There was a bid on quality, there was a bid on compounders such that they were vulnerable to any type of assumption changing. So if the growth rate declines a little bit, they get punished severely. But I remember telling you these are boring companies, that niche industries, those are the pricing power companies. And as we make a pivot away from the crowded trade and people look for safety, quality. I think one of your previous interviewees, recent episode mentioned this rotation back towards quality. Then I think this is the type of names that could see a bid. There's nothing particular about ajg. I just sent it to you as a representative. It could be in any industry. I also sent you, I think Rollins, I don't know if I mentioned Rowlands is a Terminator company. Right. Most random company, right?
- [29:45]
A
Terminator.
- [29:46]
B
Terminator like bugs you literally like in the house.
- [29:49]
A
Not like the robot Terminator.
- [29:50]
B
No, Terminator exactly like you spray Terminator like Roland's. Same thing, same chart. And then it get cut. Right. Because the growth rate declined a little bit. There's nothing wrong with the company and you should have no reason to expect that the demand for bug termination somehow declined. But it's just one of those things where valuation went, went up too much and it came down and then actually there's nothing wrong with the business. And also I think it's part of. It is also just that style of trading, the style of investing fell out of favor in a world of AI where every single cent is being pushed into the AI picks and shovel trade. And it's kind of funny like today, this morning I saw that now it's kind of almost mechanical. Like hardware goes up, software goes down, even though there's no bad news for software. And the next day is like software on the day the AI trade will get punished. And then software actually go up. Even there's nothing fundamental. So it's all very mechanical in that world. I can imagine where you can actually see a rotation back towards this type of quality. Names that are a little bit orthogonal orthogonal net beneficiary of AI adoption.
- [30:59]
A
Explain what you mean by orthogonal orthogonal
- [31:01]
B
in that the business of bug termination has nothing to do with AI doing well or not doing well. It's very, very orthogonal in that it is independent. Right. It's a different business like insurance. Same thing. Right. And those are instances where there could be a net beneficiary to AI adoption because maybe they get punished, but then next thing you know, oh, actually FDS actually became an AI adopter and they improved their product. Who knows? I don't know how easy it is for AJG to adopt insurance AI, but I suspect there's a chance.
- [31:38]
A
Right.
- [31:38]
B
But in any event, it is not something that it was being banked on as being something that need to happen. So if it did happen, it could actually be a cherry on top of a situation.
- [31:49]
A
So if I remember correctly, there's a company called Lemonade which is a AI plus Insurance, something like that.
- [31:56]
B
Yeah.
- [31:56]
A
And it had a massive run up at one point last year. It pulled back a little bit. That's one company that is kind of the new age version of this.
- [32:04]
B
Maybe see I don't like that sort of thing. Right. I, I prefer things that don't are not priced. I get it for free. Right. I like an insurance company that never had a hope of adopting such an old school €200 company and then suddenly, oh wait, this thing, not only is it not going away, but also you have to become an AI adopter. I like that. I don't like a company that comes. I mean, nothing against lemonade for lemonade. AI adoption or AI driven is already in the thesis. Right. So it had to work for that thing to work work. You know what I'm saying?
- [32:36]
A
Yes. I had another friend, he's been on this for a while. Waste management, he really likes.
- [32:43]
B
Same idea, pricing power.
- [32:45]
A
It's been an unbelievable. The chart looks insane. It's essentially just a very straight exponential curve smoothing.
- [32:52]
B
You know what else was like that? Nike. Same thing, you know. Yeah.
- [32:56]
A
Well, Nike's.
- [32:56]
B
Nike's been dogs and it's been dogs decimated.
- [32:59]
A
Yeah. And I assume Nike's probably going to try to come up with some pivot to AI at some point, try to pull an allbirds. Okay, so I think about all the AI trade and compute data that we're talking about here. How do you think something like Deepseek fits into the equation? I have a lot of friends increasingly telling me they're using Deepseek because it's cheap, it's fast, and if you're using it for everyday stuff, it's pretty much indistinguishable from some of the US models.
- [33:33]
B
I think Deep Seek is almost kind of an analogy, Right. We've been having a deep SEQ kind of a moment in AI. And more recently you have had this GLM 5.2. But if you look cross sectionally, we've also had Deep seek moments with EVs. Right. Like the reason why Elon and Tesla pivoted away from making cars towards leaning into autonomous driving and Optimus, you know, is because the Chinese EVs, right. You know, BYDs of the world, they're kind of Deep seek, right? In that they're doing kind of the same thing, but a lot cheaper. Right. At a substantially lower cost. Now, Tesla benefits from the fact that the Chinese EVs are not allowed in the US market practically. Right. They are absolutely dominating European markets. But the most valuable market, the highest paying market, the US Chinese EV is not allowed to be in. But nevertheless, it represents a threat to Tesla's global economic footprint. Same thing with AI. I think that the threat is more of a latency threat. And it goes back to this question earlier about anthropic. Anthropic faces the challenge on the top end, but on the bottom end is being chased from behind by all these open weights models that are sort of chasing and catching up in terms of capability, but at a tiny fraction of the cost. Now what they don't have is a good UX deployment in the US because they are not US based companies. And there is a general geopolitical tension. You know, this is sort of that geopolitical trade thing, right? And you know we are going to have, we have been seeing more Chinese open waste models being trained and served in the US by inference platforms, you know, US based companies, because these are open waste models, they take them here, host them here on US servers and then they do whatever modifications and then they have a certain use case. But, but I think to the extent that this is being proven out, I will be very surprised if we don't see a resurgence of so to speak, open waste models or licenseable models by other model providers that are not anthropic or OpenAI or maybe even themselves. So if you think about Meta or you think about Grok Xai, it is not inconceivable to me that these companies in fact, fact for the longest time, llama, right, that's the US Open waste models, they could provide versions that can be licensed by US companies so that you have this system where US companies are using US based open waste models and you get around this whole fear about geopolitical risks.
- [36:16]
A
So Steve, tell me. We've been using your data for this whole conversation from Silicon Data and you recently joined this company. You were previously at Bloomberg as a quant researcher and this. Yeah, this is your third time on this podcast now and we've had very different conversations each time. Tell me about what you're doing at Silicon Data now.
- [36:38]
B
So Silicon Data is a startup company that wants to bring data to the AI fiscal compute market in particular. One of the big things we're doing is we are creating so these indices, GPU rental indices, we are actually turning them into futures and we're launching futures contracts on these, on the CME in the fall. The reason we want this is because AI has become now from a strictly technology problem to a capital intensive scaling problem that is all about building data centers brick by brick and acquiring power. It's very expensive. So how do you fund this stuff? How do you actually manage the capacity? How do you make sure you're not underbuilding or overbuilding? Right. So there was that Famous moment of Dario Amode on the Cash podcast where he said the difference is between bankruptcy and not having enough. And it turns out Anthropolis didn't have enough. So we want to create futures contracts for GPU rental so that you can actually have an instrument with which you can actually hedge the risk while you are acquiring compute. Whether you are design that's actually bringing new compute onto the market in the future, or you are enterprise user of compute, essentially trying to acquire compute for your own model such that you can actually have clarity into the future. So it's not just today, but the future. So we have a forward curve of compute prices, so that's part of what we're doing. But we also one stop shop of platform of other data on AI. We have spoken about the token index that's on the consumption side of AI. But there's also we have a tool called Silicon Mark where we actually visit individual GPUs at the EU ID level, but on the serial number in the data center. You can actually go visit and get a sense of the operational efficiency and health of the GPU. So in the event of a bunch of GPUs that need to be either sold or refinanced, you can actually above beyond the rental income and that's implied by the forward curve. You can actually get a more specific set of information about the performance of this batch of GPUs. That's another thing we look at. We also have this thing called Site IQ where we give you a sense of the operational efficiency and the profitability of data centers based on the data we have so that people can form views on that as well. So it's a bunch of things.
- [39:13]
A
Wow, it's. I mean, I know you guys are at the cutting edge of this stuff and my last assumption for the day, Steve, you have chosen to work in a AI company, effectively a compute data company, that leads me to believe you are firmly in the not a bubble camp. Is that right?
- [39:36]
B
It's not like I don't think through that lens, I don't think through that phrase. I think I've been calling AI a baby bubble since 2023. So it's not like the word bubble is not in my lexicon. Right. I think asset prices can always run ahead of fundamentals. Right. And it doesn't have to invalidate the underlying story. So insofar as are we on the precipice, to me this is the only operational usefulness of the word bubble is if you actually are calling a topic top. If you are saying we've peaked and from this point forward we're going to have a big correction, a sell off like the dot com bubble bursting kind of a moment where the major participants of the trade do not recover for decades. I don't think we are in that moment. Right. Can we get to that moment at some point in the future? Possibly. Right. I think it's possible that we overbuild, but the future is not certain. It is an Observation n of 1. It's pretty interesting how you have all these quants that is basically operating on a sample size of one of like oh, there's this one set of events that occurred during the dot com bubble in this particular fashion. Then we fit every single curve onto that timeline to see where we are. I don't know whether or not it actually has to transpire that way. I do know that so to speak a bubble to me means that this entire cycle of technology adoption and build out cycle that is quite likely since we don't know what is the optimal amount and there is all sorts of FOMO that we could overbuild. And when you overbuild you can come to a moment where the fundamentals do not justify the asset value. But that could also happen even if the fundamentals do not actually ends up being overbuilt. Just like from time to time you can have a small correction because as prices run too far ahead.
- [41:31]
A
Couldn't have said it better myself. Steve, thank you very much for your time as always and you're welcome back on the show anytime.
- [41:38]
B
Thank you so much for having me.